LLM writing workflow emphasizes critique over ghostwriting to preserve author voice

The source story does not document a specific deployed model, benchmark, architecture, release channel, or API version. Ptacek mentions using “a good model,” later names GPT5 in an anecdotal proofreading example, and suggests a local workshopping tool could run prompts through Codex, Claude, or Antigravity CLIs. No exact model snapshot, provider settings, temperature, system prompt, context length, or evaluation harness is specified. How To Write With An LLM — A Final Ward

The only concrete implementation stack in the essay is for a proposed writing-workshopping application: Python, HTMX, SQLite, and Tailwind, with a Notion-style editor, highlighted passages, sidebar commentary, revision tracking, and navigation through suggestions. That is a product/workflow sketch, not a reproducible software release. How To Write With An LLM — A Final Ward

Separately, Google describes Antigravity CLI as bringing the Antigravity agent harness into a terminal, but Ptacek’s post does not depend on Google’s product claims or document a specific Antigravity configuration. GitHub - google-antigravity/antigravity-cli: Antigravity CLI brings the reasoning, execution, and orchestration capabilities of Antigravity agent harness directly into your terminal. · GitHub

Architecture: what is documented

The documented architecture is a human-authored draft plus LLM critique loop:

1. Human writes the original draft.
2. LLM flags problems: repetition, passive voice, nominalizations, weak flow, movable paragraphs, filler adverbs, grammar, typos, logic, or factual issues.
3. Human rewrites the relevant passage in their own words.
4. Optionally, an LLM compares original vs. revised passages, ideally without being told which one the author just produced.
5. The human accepts or rejects the advice.

This is closer to a diagnostic assistant than a generative writing system. The key design choice is that model output is treated as an annotation layer, not a source of final prose. Ptacek’s own tool concept uses highlighted spans and sidebar comments, which mirrors code-review UI more than conventional chat. How To Write With An LLM — A Final Ward

Evaluation methodology: what is missing

There is no controlled evaluation in the source article. It does not measure time saved, edit quality, reader preference, factual accuracy, stylistic preservation, or brand lift. It also does not compare models. Any claim that this workflow is “better” should therefore be read as practitioner judgment, not independently measured performance. How To Write With An LLM — A Final Ward

A reproducible evaluation would need at least: paired drafts, blind human raters, task categories, author satisfaction measures, voice-preservation measures, factual-error audits, and a condition comparing ghostwriting, copyediting, grammar-only tools, and human-only editing. The PNAS study on LLM style shows that corpus-linguistic features and source-classification methods can detect systematic differences between human and model prose, suggesting a possible measurement approach for “voice drift.” Do LLMs write like humans? Variation in grammatical and rhetorical styles | PNAS

Implementation implications

For product teams building AI writing tools, the workflow implies a different UI contract:

• Prefer comments, highlights, and checklists over full rewritten paragraphs.
• Separate diagnostic suggestions from replacement prose.
• Provide a “no praise / no encouragement” mode.
• Log which passages were model-flagged and which were human-rewritten.
• Support blind A/B review, but do not treat model preference as ground truth.
• For fact-checking, require citations or source links and human verification.

Willison’s proofreading prompt is a concrete lightweight version: it asks the model to identify spelling, grammar, repetition, logical errors, factual mistakes, weak arguments, and placeholder links, while his policy says opinionated first-person writing must be his own. Prompts I use - Agentic Engineering Patterns - Simon Willison's Weblog

Ptacek’s argument fits a broader move from “AI writes for you” toward AI as critic, reviewer, or Socratic counterpart. In education and writing research, some scholars argue that writing with LLMs should be taught as a dialogical practice rather than simple text outsourcing; the key question becomes who retains rhetorical agency and responsibility. Writing with ChatGPT

The empirical backdrop is mixed. Research on creative ideation suggests LLMs can make individuals feel more productive or produce more detailed outputs, while reducing distinctiveness across users. 2402.01536 Homogenization Effects of Large Language Models on Human Creative Ideation Research on culturally grounded writing found that GPT-4o autocomplete suggestions changed how Indian participants wrote, pushing style toward Western norms. 2409.11360 AI Suggestions Homogenize Writing Toward Western Styles and Diminish Cultural Nuances And the PNAS work on rhetorical style found that instruction-tuned models can produce a dense, noun-heavy style and struggle to match genre variation even when prompted to imitate human text. Do LLMs write like humans? Variation in grammatical and rhetorical styles | PNAS

These studies do not validate Ptacek’s exact rules. But they make his concern plausible: if models have default stylistic attractors, then accepting model phrasings can create a convergence pressure on voice.

The strongest limitation is that the source is commentary, not a controlled experiment. Ptacek’s rules are disciplined craft advice from an experienced practitioner; they are not a measured safety standard. How To Write With An LLM — A Final Ward

There is also disagreement about strictness. Bruce R. Lewis agrees with using LLMs as copyeditors but argues that banning even a single word is too strong, because a model can serve as a more flexible thesaurus than conventional synonym lookup. LLM as copyeditor - Writing by Bruce R. Lewis (brlewis) Hacker News commenters similarly split between “don’t use it for writing,” “proofreading is enough,” and “models can be useful as reader-role simulators or fact-checking prompts.” How to Write with an LLM | Hacker News

Fact-checking is the highest-risk part of the workflow. A model can identify possible factual issues, but it can also hallucinate. OpenAI’s help center explicitly states that ChatGPT may produce factually inaccurate responses and recommends using tools like search or deep research and checking sources directly when accuracy matters. Does ChatGPT tell the truth? | OpenAI Help Center

For executives, this story is about workflow design and accountability, not writing-tool procurement. Teams that publish thought leadership, developer relations content, policy material, legal-adjacent guidance, or executive communications should define which uses are allowed:

• Allowed: typo detection, grammar review, repetition detection, argument stress-testing, source checking, empty-link detection.
• Restricted: wholesale rewrites, undisclosed ghostwriting, unsupported factual claims, model-generated executive voice.
• Auditable: final human owner, source list, model role, and whether any model-generated sentence entered final copy.

For developers, the opportunity is to build writing tools that act like linters for prose: flag issues, localize evidence, avoid praise, and leave the final patch to the author. That design aligns better with preserving authorship than chatbots that return polished replacement paragraphs.